AI Can Now Work for 12 Hours Straight: The Self-Improving AI Era Has Begun

A new blog post by Anthropic co-founder Jack Clark has sent ripples through the AI research community this week. Clark put 60%+ odds on AI systems training their own successors before 2029 — not as speculative fiction, but as a data-driven forecast based on publicly available benchmarks showing AI capability compounding at rates that most observers have not fully registered.

The Numbers Behind the Forecast

Clark's case is built on concrete benchmark data rather than theoretical projections. In 2022, AI could handle autonomous tasks lasting roughly 30 seconds before requiring human intervention. By 2026, that autonomous working window has extended to 12 hours — meaning AI systems can execute complex, multi-step tasks independently for half a working day without human input. Projections based on the current improvement curve suggest 100-hour autonomous runs by the end of 2026.

On the SWE-Bench benchmark — which tests AI on real GitHub coding tasks, the kind of work that professional software engineers do daily — the progression has been extraordinary. Claude 2, released less than three years ago, scored just 2%. Anthropic's Mythos Preview has now scored 93.9% on the same benchmark. That is not incremental improvement. It is a near-complete capability acquisition in under three years.

What Does Self-Improving AI Actually Mean?

Self-improving AI refers to systems that can meaningfully contribute to the development of their own successors — writing code, running experiments, analysing results, and proposing architectural improvements that human researchers then evaluate and implement. It does not require the AI to be fully autonomous in the science of AI development. It requires only that AI becomes sufficiently capable at the core tasks of AI research — coding, hypothesis generation, experiment design, and result analysis — to accelerate the development cycle significantly.

Given that Mythos Preview scores 93.9% on real software engineering tasks, the question of whether AI can contribute to AI development is no longer theoretical. The question is at what scale and with what level of human oversight that contribution is happening right now, and how quickly it will become the dominant driver of capability improvement.

OpenAI's Automated Research Intern by September 2026

Clark's forecast aligns with concrete targets being set inside the leading AI labs. OpenAI has internally targeted the deployment of an 'automated research intern' — an AI system capable of performing the work of a junior AI researcher — by September 2026. This is not a moonshot goal set by optimistic executives. It is a near-term operational target based on current capability trajectories. METR, an AI evaluation organisation, has independently tracked AI autonomous task completion time doubling approximately every six months — a pace that, if sustained, makes 100-hour autonomous research runs a 2026 milestone rather than a 2030 prediction.

Why This Matters More Than Most AI News

Most AI news covers what AI can do today — which models write the best code, which chatbot gives the most accurate answers. The self-improving AI story is different because it is about what determines the pace of all future AI development. If AI systems become significant contributors to their own improvement, the development cycle compresses from years to months, or months to weeks. Human researchers set the direction, safety constraints, and evaluation criteria — but the execution of AI research becomes increasingly AI-driven.

This is the context in which Anthropic's Mythos, the White House's regulatory concerns, and the massive infrastructure investments being made across the industry make complete sense. Everyone working seriously on AI understands that the capability curve is compounding, not linear. The decisions being made right now — about safety frameworks, compute access, regulatory oversight, and who controls the frontier models — will shape what the self-improving AI era looks like when it fully arrives.

What Should You Do With This Information?

For individuals, the self-improving AI trajectory reinforces one clear priority: develop skills that compound with AI rather than compete against it. Understanding how to direct AI systems, evaluate their output, and integrate them into complex workflows is more valuable now than almost any technical skill that AI itself is rapidly acquiring. For businesses, the timeline suggests that AI capabilities available in 2028 will be dramatically different from those available today — and planning for that shift now, rather than reacting to it, is the competitive advantage.